📚 Ask-My-Documents

A production-grade Multimodal RAG (Retrieval-Augmented Generation) system for intelligent document understanding, semantic retrieval, and grounded question answering.

---

Overview

Ask-My-Documents is an advanced multimodal RAG pipeline designed to process complex documents such as research papers, technical PDFs, and enterprise documents. Unlike basic PDF chatbot implementations, this project focuses on retrieval quality, multimodal understanding, and production-style document ingestion pipelines.

The system combines:

• Semantic chunking — title-aware, context-preserving document segmentation
• OCR-aware extraction — robust handling of scanned and image-heavy documents
• Table-aware parsing — structured preservation of tabular data
• Image-aware understanding — vision-language model integration for multimodal content
• AI-powered semantic enrichment — LLM-augmented chunk metadata
• Vector retrieval pipelines — high-performance semantic search via ChromaDB

---

Features

Feature	Description
📄 PDF Ingestion	Parse and extract content from complex PDFs
🧠 Semantic Enrichment	AI-enhanced chunk-level understanding
🖼️ Multimodal Understanding	Vision-Language Model integration (Qwen2-VL)
📊 Table Extraction	Structure-preserving table parsing
🔎 Vector Retrieval	Semantic search with ChromaDB
🧩 Title-Aware Chunking	Context-coherent document segmentation
📚 Grounded Generation	Source-cited answer generation
🏷️ Metadata-Aware Retrieval	Rich metadata for precision filtering
🚀 Production Architecture	Modular, extensible ingestion pipeline

---

Tech Stack

Component	Technology
Document Parsing	Unstructured
Embeddings	BAAI/bge-small-en-v1.5
Vector Database	ChromaDB
Multimodal Model	Qwen2-VL
Framework	LangChain
OCR	Tesseract OCR
PDF Processing	Poppler
Backend	Python 3.10+

---

System Architecture

Document Upload
      │
      ▼
Document Parsing (Unstructured)
      │
      ▼
Semantic Element Extraction
      │
      ▼
Title-Aware Chunking
      │
      ▼
Multimodal Content Extraction
      │
      ▼
AI Semantic Enrichment
      │
      ▼
Embeddings Generation
      │
      ▼
ChromaDB Vector Storage
      │
      ▼
Hybrid Retrieval + Re-ranking
      │
      ▼
Grounded Answer Generation

---

Project Structure

ask-my-documents/
│
├── data/
├── notebooks/
├── src/
│   ├── ingestion/
│   ├── chunking/
│   ├── embeddings/
│   ├── retrieval/
│   ├── llm/
│   ├── evaluation/
│   └── utils/
│
├── vector_db/
├── app/
├── requirements.txt
└── README.md

---

Installation

1. Clone the Repository

git clone https://github.com/your-username/ask-my-documents.git
cd ask-my-documents

2. Create a Virtual Environment

python -m venv venv

# Windows
venv\Scripts\activate

# Linux / macOS
source venv/bin/activate

3. Install Python Dependencies

pip install -r requirements.txt

4. Install System Dependencies

Linux:

sudo apt-get install poppler-utils tesseract-ocr

Windows:

Download and install:

• Poppler for Windows
• Tesseract OCR

Then add both to your system PATH.

---

Current Capabilities

• Semantic document chunking
• Multimodal preprocessing pipeline
• OCR-aware extraction
• Table-aware document understanding
• Image-aware retrieval enrichment
• Vision-language model integration
• Retrieval-ready semantic indexing

---

Roadmap

• Hybrid Retrieval (BM25 + Dense Retrieval)
• Cross-Encoder Re-ranking
• RAG Evaluation Pipeline (Ragas)
• Citation-aware responses
• Streamlit / FastAPI deployment
• Image embedding retrieval
• Parent-child retrieval
• Cross-modal search

---

Example Use Cases

• 📖 Research Paper Assistant — Query academic papers with grounded citations
• 🛠️ Technical Documentation QA — Instant answers from complex technical manuals
• 🏢 Enterprise Document Search — Retrieve insights from internal knowledge bases
• 📊 Table-Aware QA — Ask questions directly about tabular data
• 🌐 Multimodal Knowledge Retrieval — Combine text and image understanding

---

License

This project is licensed under the MIT License.

---

Author

Developed as part of an advanced RAG engineering and multimodal retrieval learning journey.